
Buying a GEO audit: questions to ask before approving the proposal
Find out whether an AI-search audit will give you useful evidence, a clear action plan and something your team can repeat.
Written by Rob KnottA colourful visibility score can look reassuring without telling you what your business should do next. Before approving a GEO audit, ask what decision it will help you make and what evidence you will receive.
GEO means generative engine optimisation. In this context, an audit investigates how your business appears in AI-generated answers and what may help improve that representation. Suppliers package this work differently, so compare the scope and method rather than relying on the label.
Start with the decision you need to make
You might need to know whether your brand enters a buyer's shortlist, whether a product is described accurately, which competitors appear in a target market or where your content is being cited. Those questions require different samples and evidence.
Write a short commissioning statement: “We want to understand how our service is represented during supplier comparison in the UK, identify important inaccuracies and agree the first changes to investigate.” That is more useful than “We want a GEO score”.
Define the audience, products, markets, languages and search environments. Include what you will do with the findings. A team able to improve product documentation needs different recommendations from one commissioning a major technical rebuild.
Ask how the questions are selected
A supplier can produce a flattering result by asking questions that name your brand or describe it unusually closely. Branded prompts can be useful for checking accuracy, but they should not stand in for discovery research.
Ask for the rationale behind the sample. Are questions informed by customer conversations, sales objections, search data or product journeys? Which are branded, unbranded, comparative or informational? How are markets and languages represented? You should be able to see what was included and what was left out.
More prompts do not automatically mean more representative coverage. Hundreds of near-duplicates can overwhelm a report while missing a critical buying decision. Ask the supplier to show how the sample maps to your business questions and how changes to that sample will be documented.
Separate the measures before comparing scores
A brand mention is different from a recommendation. A citation to your website is different from a citation to a review site that discusses you. An answer can name you prominently while giving incorrect information.
Require definitions for each reported measure and a route back to the observation. If the supplier combines them into a score, ask for the formula, component values and treatment of missing or failed observations. A score change should be explainable.
Native reporting and observed answers also have different coverage. Bing's reporting describes citations within the experiences it covers; it is not a census of every assistant. Google's guidance describes its own AI features. Ask suppliers to state the scope of each data source instead of blending everything into an unqualified “AI market share”.
| Measure | What it can tell you | What it does not establish |
|---|---|---|
| Brand presence | Whether the brand appears in recorded answers | That a person saw or trusted the answer |
| Recommendation | Whether the answer presents the brand as an option | A sale or qualified enquiry |
| Citation | Whether a source is visibly referenced | That the cited source is the only influence |
| Accuracy | Whether specified claims match verified facts | That all possible claims have been checked |
| Referral visits | Visits attributed to recognised sources | Every AI-influenced customer journey |
| Business outcomes | Recorded enquiries or sales and their context | A complete causal attribution to the audit |
Look for a repeatable collection method
Ask when observations were collected, which platform or model was visible, whether search was enabled, and what session, location or language settings were used. Some controls cannot be perfectly standardised. The report should record those limits rather than conceal them.
Repeated observations help reveal variation. There is no universal number of runs that makes every audit statistically reliable. A modest exploratory sample can still be useful if its limitations are clear. It becomes misleading when a small convenience sample is presented as a precise estimate of all customer behaviour.
Ask how failed runs, unavailable answers and platform changes are handled. A missing response is not automatically a zero for your brand. You also need to know whether a later comparison uses the same core sample and collection conditions.
For the detailed research method, read the AI visibility measurement framework.
Worked example: comparing two fictional proposals
Suppose a B2B software business receives two proposals. Proposal A offers 500 prompts and a single headline score. Proposal B offers 60 questions grouped around discovery, comparison and validation, repeated observations, an evidence register and a workshop to prioritise findings.
The smaller number does not automatically make B better. The buyer asks A for its sampling logic and underlying evidence, and asks B whether the 60 questions cover the important markets and use cases. Both are judged against the actual decision, not an impressive count.
If A can explain its method and provide usable evidence, it may be suitable. If B cannot cover a commercially important audience, its attractive methodology still needs revision. The useful buying behaviour is asking for the missing detail before signing.
Make recommendations traceable and practical
A finding should distinguish what was observed, the possible explanation and the proposed action. Seeing a competitor cited does not prove why it was selected. A recommendation to improve an important factual page may be sensible, but the report should not pretend that the observation establishes a secret ranking rule.
Useful actions identify the owner, expected effort, dependencies and how to check implementation. Separate technical faults, factual corrections, content opportunities and offsite representation. Your team should be able to act on the priorities without reconstructing the analyst's reasoning.
Ask what happens after delivery. Does the price include a clarification session, a prioritisation workshop or a repeat baseline? Does anyone implement changes? An audit and an ongoing service can both be valuable, but their boundaries should be explicit.
The AI visibility audit guide explains how findings become investigative questions and actions.
Agree ownership and acceptance before purchase
Ensure your business receives the agreed outputs in a usable format. That may include the prompt list, definitions, permitted evidence exports, finding register, limitations and action plan. Clarify licences and confidentiality where third-party tools or platform terms constrain what can be handed over.
Define completion by the agreed work being delivered and explained. Avoid acceptance criteria that require a guaranteed increase in recommendations or citations. Those outcomes depend on systems outside the supplier's control.
Use the questions below in your proposal review. “Not included” is an acceptable answer when you understand the consequence and have another way to cover the need.
- Which business decisions, audiences and markets does this audit cover?
- Can we inspect the prompt sample and its selection rationale?
- Which measures are reported separately, with what definitions?
- Can each important finding be traced to dated evidence?
- How are variation, failed observations and tool coverage documented?
- Will recommendations identify owners, dependencies and verification?
- What files and usage rights do we receive?
- What interpretation, implementation or repeat measurement is included?
Commission the work around your next decision
If you are unsure what to ask for, start with a short scoping discussion rather than buying the largest package. Establish the questions that matter, the evidence already available and the actions your team could take.
A well-scoped baseline should help you understand your position and choose a credible next step. It can be a defined project with a handover, or the start of further advisory support if your team needs help applying it.
Key takeaways
- Buy evidence and decision support rather than an unexplained score.
- Inspect sampling, definitions, repeatability and platform coverage.
- Request a worked deliverable before comparing prompt counts.
- Agree what your team receives and what follow-up support includes.

